(简体中文|English)
随着互联网不断发展,电子邮件、社交媒体照片、直播视频、客服语音等非结构化数据已经变得越来越普遍。如果想要使用计算机来处理这些数据,需要使用 embedding 技术将这些数据转化为向量 vector,然后进行存储、建索引、并查询。
但是,当数据量很大,比如上亿条音频要做相似度搜索,就比较困难了。穷举法固然可行,但非常耗时。针对这种场景,该 demo 将介绍如何使用开源向量数据库 Milvus 搭建音频相似度检索系统。
音频检索(如演讲、音乐、说话人等检索)实现了在海量音频数据中查询并找出相似声音(或相同说话人)片段。音频相似性检索系统可用于识别相似的音效、最大限度减少知识产权侵权等,还可以快速的检索声纹库、帮助企业控制欺诈和身份盗用等。在音频数据的分类和统计分析中,音频检索也发挥着重要作用。
在本 demo 中,你将学会如何构建一个音频检索系统,用来检索相似的声音片段。使用基于 PaddleSpeech 预训练模型(音频分类模型,说话人识别模型等)将上传的音频片段转换为向量数据,并存储在 Milvus 中。Milvus 自动为每个向量生成唯一的 ID,然后将 ID 和 相应的音频信息(音频id,音频的说话人id等等)存储在 MySQL,这样就完成建库的工作。用户在检索时,上传测试音频,得到向量,然后在 Milvus 中进行向量相似度搜索,Milvus 返回的检索结果为向量 ID,通过 ID 在 MySQL 内部查询相应的音频信息即可。
注:该 demo 使用 CN-Celeb 数据集,包括至少 650000 条音频,3000 个说话人,来建立音频向量库(音频特征,或音频说话人特征),然后通过预设的距离计算方式进行音频(或说话人)检索,这里面数据集也可以使用其他的,根据需要调整,如Librispeech,VoxCeleb,UrbanSound,GloVe,MNIST等。
音频向量的提取需要用到基于 PaddleSpeech 训练的模型,所以请确保在运行之前已经安装了 PaddleSpeech,具体安装步骤,详见安装文档。
你可以从 easy,medium,hard 三种方式中选择一种方式安装。
音频相似性的检索需要用到 Milvus, MySQL 服务。 我们可以通过 docker-compose.yaml 一键启动这些容器,所以请确保在运行之前已经安装了 Docker Engine 和 Docker Compose。 即
## 先进入到 audio_searching 目录,如下示例
cd ~/PaddleSpeech/demos/audio_searching/
## 然后启动容器内的相关服务
docker-compose -f docker-compose.yaml up -d
你会看到所有的容器都被创建:
Creating network "quick_deploy_app_net" with driver "bridge"
Creating milvus-minio ... done
Creating milvus-etcd ... done
Creating audio-mysql ... done
Creating milvus-standalone ... done
Creating audio-webclient ... done
可以采用'docker ps'来显示所有的容器,还可以使用'docker logs audio-mysql'来获取服务器容器的日志:
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
b2bcf279e599 milvusdb/milvus:v2.0.1 "/tini -- milvus run…" 22 hours ago Up 22 hours 0.0.0.0:19530->19530/tcp milvus-standalone
d8ef4c84e25c mysql:5.7 "docker-entrypoint.s…" 22 hours ago Up 22 hours 0.0.0.0:3306->3306/tcp, 33060/tcp audio-mysql
8fb501edb4f3 quay.io/coreos/etcd:v3.5.0 "etcd -advertise-cli…" 22 hours ago Up 22 hours 2379-2380/tcp milvus-etcd
ffce340b3790 minio/minio:RELEASE.2020-12-03T00-03-10Z "/usr/bin/docker-ent…" 22 hours ago Up 22 hours (healthy) 9000/tcp milvus-minio
15c84a506754 paddlepaddle/paddlespeech-audio-search-client:2.3 "/bin/bash -c '/usr/…" 22 hours ago Up 22 hours (healthy) 0.0.0.0:8068->80/tcp audio-webclient
启动系统服务程序,它会提供基于 HTTP 后端服务。
-
安装服务依赖的 python 基础包
pip install -r requirements.txt
-
修改配置(本地运行情况下,一般不用修改,可以跳过该步骤)
## 方法一:修改源码文件 vim src/config.py ## 方法二:修改环境变量,如下所示 export MILVUS_HOST=127.0.0.1 export MYSQL_HOST=127.0.0.1
这里列出了一些需要设置的参数,更多信息请参考 config.py
参数 描述 默认设置 MILVUS_HOST Milvus 服务的 IP 地址 127.0.0.1 MILVUS_PORT Milvus 服务的端口号 19530 VECTOR_DIMENSION 特征向量的维度 192 MYSQL_HOST Mysql 服务的 IP 地址 127.0.0.1 MYSQL_PORT Mysql 服务的端口号 3306 DEFAULT_TABLE 默认存储的表名 audio_table -
运行程序
启动用 Fastapi 构建的服务
export PYTHONPATH=$PYTHONPATH:./src python src/audio_search.py
然后你会看到应用程序启动:
INFO: Started server process [13352] 2022-03-26 22:45:30,838 | INFO | server.py | serve | 75 | Started server process [13352] INFO: Waiting for application startup. 2022-03-26 22:45:30,839 | INFO | on.py | startup | 45 | Waiting for application startup. INFO: Application startup complete. 2022-03-26 22:45:30,839 | INFO | on.py | startup | 59 | Application startup complete. INFO: Uvicorn running on http://0.0.0.0:8002 (Press CTRL+C to quit) 2022-03-26 22:45:30,840 | INFO | server.py | _log_started_message | 206 | Uvicorn running on http://0.0.0.0:8002 (Press CTRL+C to quit)
-
准备数据
wget -c https://www.openslr.org/resources/82/cn-celeb_v2.tar.gz && tar -xvf cn-celeb_v2.tar.gz
注:如果希望快速搭建 demo,可以采用 ./src/test_audio_search.py:download_audio_data 内部的 20 条音频,另外后续结果展示以该集合为例
-
准备模型(如果使用默认模型,可以跳过此步骤)
## 修改模型配置参数,目前 model 仅支持 ecapatdnn_voxceleb12,后续将支持多种类型 vim ./src/encode.py
-
脚本测试(推荐)
python ./src/test_audio_search.py
注:内部将依次下载数据,加载 paddlespeech 模型,提取 embedding,存储建库,检索,删库
输出:
Downloading https://paddlespeech.bj.bcebos.com/vector/audio/example_audio.tar.gz ... ... Unpacking ./example_audio.tar.gz ... [2022-03-26 22:50:54,987] [ INFO] - checking the aduio file format...... [2022-03-26 22:50:54,987] [ INFO] - The sample rate is 16000 [2022-03-26 22:50:54,987] [ INFO] - The audio file format is right [2022-03-26 22:50:54,988] [ INFO] - device type: cpu [2022-03-26 22:50:54,988] [ INFO] - load the pretrained model: ecapatdnn_voxceleb12-16k [2022-03-26 22:50:54,990] [ INFO] - Downloading sv0_ecapa_tdnn_voxceleb12_ckpt_0_1_0.tar.gz from https://paddlespeech.bj.bcebos.com/vector/voxceleb/sv0_ecapa_tdnn_voxceleb12_ckpt_0_1_0.tar.gz ... [2022-03-26 22:51:17,285] [ INFO] - start to dynamic import the model class [2022-03-26 22:51:17,285] [ INFO] - model name ecapatdnn [2022-03-26 22:51:23,864] [ INFO] - start to set the model parameters to model [2022-03-26 22:54:08,115] [ INFO] - create the model instance success [2022-03-26 22:54:08,116] [ INFO] - Preprocess audio file: /home/zhaoqingen/PaddleSpeech/demos/audio_ searching/example_audio/knife_hit_iron3.wav [2022-03-26 22:54:08,116] [ INFO] - load the audio sample points, shape is: (11012,) [2022-03-26 22:54:08,150] [ INFO] - extract the audio feat, shape is: (80, 69) [2022-03-26 22:54:08,152] [ INFO] - feats shape: [1, 80, 69] [2022-03-26 22:54:08,154] [ INFO] - audio extract the feat success [2022-03-26 22:54:08,155] [ INFO] - start to do backbone network model forward [2022-03-26 22:54:08,155] [ INFO] - feats shape:[1, 80, 69], lengths shape: [1] [2022-03-26 22:54:08,433] [ INFO] - embedding size: (192,) Extracting feature from audio No. 1 , 20 audios in total [2022-03-26 22:54:08,435] [ INFO] - checking the aduio file format...... [2022-03-26 22:54:08,435] [ INFO] - The sample rate is 16000 [2022-03-26 22:54:08,436] [ INFO] - The audio file format is right [2022-03-26 22:54:08,436] [ INFO] - device type: cpu [2022-03-26 22:54:08,436] [ INFO] - Model has been initialized [2022-03-26 22:54:08,436] [ INFO] - Preprocess audio file: /home/zhaoqingen/PaddleSpeech/demos/audio_searching/example_audio/sword_wielding.wav [2022-03-26 22:54:08,436] [ INFO] - load the audio sample points, shape is: (6391,) [2022-03-26 22:54:08,452] [ INFO] - extract the audio feat, shape is: (80, 40) [2022-03-26 22:54:08,454] [ INFO] - feats shape: [1, 80, 40] [2022-03-26 22:54:08,454] [ INFO] - audio extract the feat success [2022-03-26 22:54:08,454] [ INFO] - start to do backbone network model forward [2022-03-26 22:54:08,455] [ INFO] - feats shape:[1, 80, 40], lengths shape: [1] [2022-03-26 22:54:08,633] [ INFO] - embedding size: (192,) Extracting feature from audio No. 2 , 20 audios in total ... 2022-03-26 22:54:15,892 | INFO | audio_search.py | load_audios | 85 | Successfully loaded data, total count: 20 2022-03-26 22:54:15,908 | INFO | audio_search.py | count_audio | 148 | Successfully count the number of data! [2022-03-26 22:54:15,916] [ INFO] - checking the aduio file format...... [2022-03-26 22:54:15,916] [ INFO] - The sample rate is 16000 [2022-03-26 22:54:15,916] [ INFO] - The audio file format is right [2022-03-26 22:54:15,916] [ INFO] - device type: cpu [2022-03-26 22:54:15,916] [ INFO] - Model has been initialized [2022-03-26 22:54:15,916] [ INFO] - Preprocess audio file: /home/zhaoqingen/PaddleSpeech/demos/audio_searching/example_audio/test.wav [2022-03-26 22:54:15,917] [ INFO] - load the audio sample points, shape is: (8456,) [2022-03-26 22:54:15,923] [ INFO] - extract the audio feat, shape is: (80, 53) [2022-03-26 22:54:15,924] [ INFO] - feats shape: [1, 80, 53] [2022-03-26 22:54:15,924] [ INFO] - audio extract the feat success [2022-03-26 22:54:15,924] [ INFO] - start to do backbone network model forward [2022-03-26 22:54:15,924] [ INFO] - feats shape:[1, 80, 53], lengths shape: [1] [2022-03-26 22:54:16,051] [ INFO] - embedding size: (192,) ... 2022-03-26 22:54:16,086 | INFO | audio_search.py | search_local_audio | 132 | search result http://testserver/data?audio_path=./example_audio/test.wav, score 100.0 2022-03-26 22:54:16,087 | INFO | audio_search.py | search_local_audio | 132 | search result http://testserver/data?audio_path=./example_audio/knife_chopping.wav, score 29.182177782058716 2022-03-26 22:54:16,087 | INFO | audio_search.py | search_local_audio | 132 | search result http://testserver/data?audio_path=./example_audio/knife_cut_into_body.wav, score 22.73637056350708 ... 2022-03-26 22:54:16,088 | INFO | audio_search.py | search_local_audio | 136 | Successfully searched similar audio! 2022-03-26 22:54:17,164 | INFO | audio_search.py | drop_tables | 160 | Successfully drop tables in Milvus and MySQL!
-
前端测试(可选)
在浏览器中输入 127.0.0.1:8068 访问前端页面
注:如果浏览器和服务不在同一台机器上,那么 IP 需要修改成服务所在的机器 IP,并且 docker-compose.yaml 中相应的 API_URL 也要修改,然后重新执行 docker-compose.yaml 文件,使修改生效。
机器配置:
- 操作系统: CentOS release 7.6
- 内核:4.17.11-1.el7.elrepo.x86_64
- 处理器:Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz
- 内存:132G
数据集:
- CN-Celeb, 训练集 65万, 测试集 1万,向量维度 192,距离计算方式 L2
召回和耗时统计如下图:
基于 Milvus 的检索框架在召回率 90% 的前提下,检索耗时约 2.9 毫秒,加上特征提取(Embedding)耗时约 500 毫秒(测试音频时长约 5 秒),即单条音频测试总共耗时约 503 毫秒,可以满足大多数应用场景。
以下是 PaddleSpeech 提供的预训练模型列表:
模型 | 采样率 |
---|---|
ecapa_tdnn | 16000 |